def train_new(path_model):
gpu_id = (1 if torch.cuda.is_available() else -1)
x, y = get_data(gpu_id)
model = Model(
nn_class='test_fnet_model.DummyModel',
nn_kwargs={'some_param': SOME_PARAM_TEST_VAL},
)
model.to_gpu(gpu_id)
for idx in range(4):
_ = model.train_on_batch(x, y)
model.save(path_model)
def train_new(path_model):
gpu_id = 1 if torch.cuda.is_available() else -1
x, y = get_data(gpu_id)
model = Model(
nn_class="fnet.nn_modules.dummy.DummyModel",
nn_kwargs={"some_param": SOME_PARAM_TEST_VAL},
)
model.to_gpu(gpu_id)
for idx in range(4):
_ = model.train_on_batch(x, y)
model.save(path_model)
def test_train_on_batch():
model = Model(nn_class='tests.data.nn_test.Net', lr=0.01)
shape_item = (1, 2, 4, 8)
batch_size = 9
shape_batch = (batch_size, ) + shape_item
x_batch = torch.rand(shape_batch)
y_batch = x_batch * 0.666 + 0.42
cost_prev = float('inf')
for _ in range(8):
cost = model.train_on_batch(x_batch, y_batch)
assert cost < cost_prev
cost_prev = cost
# Test target weight maps
model = Model(nn_class='tests.data.nn_test.Net', lr=0.) # disable learning
cost_norm = model.train_on_batch(x_batch, y_batch)
# Test uniform weight map
weight_map_batch = ((torch.ones(shape_item) /
np.prod(shape_item)).expand(shape_batch))
cost_weighted = model.train_on_batch(x_batch, y_batch, weight_map_batch)
npt.assert_approx_equal(cost_weighted, cost_norm)
# Test all-zero weight map
cost_weighted = model.train_on_batch(x_batch, y_batch,
torch.zeros(x_batch.size()))
npt.assert_approx_equal(cost_weighted, 0.)
# Random weights with first and last examples having zero weight
weight_map_batch = torch.rand(shape_batch)
weight_map_batch[[0, -1]] = 0.
cost_weighted = model.train_on_batch(x_batch, y_batch, weight_map_batch)
cost_exp = (
model.train_on_batch(x_batch[1:-1], y_batch[1:-1],
weight_map_batch[1:-1]) * (batch_size - 2) /
batch_size # account for change in batch size
)
npt.assert_approx_equal(cost_weighted, cost_exp)